MiniMax H3 Shotboard Settings → CineLinX
All the MiniMax H3 Shotboard settings, exported as one CineLinX config
- cine_linx
The IAMCCS_MiniMaxH3ShotPlanner (the Shotboard) has so many widgets that splitting the config from the planning is the obvious quality-of-life move. IAMCCS_ShotboardH3Settings is that split: it's the same settings surface as the Shotboard - same contracts, same option lists - but its only job is to package them into one cine_linx output that the Shotboard (or another consumer) can read. You can't get a cleaner separation of "configure once" from "plan each shot."
How it works
Rather than maintaining a second fragile widget schema, the node literally borrows the Shotboard's input definitions (IAMCCS_MiniMaxH3ShotPlanner.INPUT_TYPES()), groups them by category, and re-exposes them with display names like 08 · ADVANCED COMPATIBILITY · ltx round mode. That means every future option the Shotboard gains - samplers, installed LoRAs, new presets - appears here automatically. It also strips the control_after_generate auto-seed behavior, because this is a persisted configuration source, not a sampler; the Shotboard keeps the single seed control.
Practically, you get the full set of Shotboard knobs here: duration_seconds, width/height (and the image-reference variants image_width/image_height with their resize method), guide_policy, default_force, promptrelay_epsilon, ltx_round_mode, and the rest of the H3 settings groups. You set them once, wire the cine_linx output into the Shotboard's cine_linx input, and the shot inherits them.
Outputs
A single output: cine_linx, carrying the settings payload under an iamccs_minimax_h3_settings schema with schema_version: 1. The node's policy strings make the priority explicit: minimax_h3_settings_source: IAMCCS_ShotboardH3Settings and minimax_h3_settings_priority: external_cine_linx_over_local_snapshot - i.e. when both an external settings CineLinX and the Shotboard's own widgets are present, the external config wins. That's the whole contract in two lines: this node is meant to be the authority, so don't fight it by also hand-editing the same fields in the Shotboard.
Why you'd bother
Two workflows make it worth it. First, presets: build two or three of these as saved config sources (a "hero quality" one, a "low VRAM draft" one) and switch between them by rewiring a single connection instead of scrolling fifty widgets. Second, consistency across many shots: if you always render the same canvas size, guide policy and round mode, locking them here removes a whole class of "one segment rendered at the wrong resolution" accidents.
Ships with IAMCCS-nodes: ComfyUI Manager → search "IAMCCS", or cd ComfyUI/custom_nodes && git clone https://github.com/IAMCCS/IAMCCS-nodes.git, then restart. No extra models or dependencies beyond the pack.
Inputs (253)
| Name | Type | Default | Description |
|---|---|---|---|
| duration_seconds | FLOAT | 10.000.01–36000 | — |
| frame_rate | INT | 2424–24 | — |
| guide_policy | COMBO | every_checked_row | 1 options: every_checked_row |
| min_guide_gap_seconds | FLOAT | 0.000–0 | — |
| max_guides | INT | 5050–50 | — |
| default_force | FLOAT | 1.001–1 | — |
| promptrelay_epsilon | FLOAT | 0.00100.001–0.001 | — |
| ltx_round_mode | COMBO | none | 1 options: none |
| width | INT | 960256–5760 | — |
| height | INT | 544256–5760 | — |
| image_width | INT | 960256–5760 | — |
| image_height | INT | 544256–5760 | — |
| image_resize_method | COMBO | crop | 5 options: crop, pad, keep proportion, stretch, |
| image_multiple_of | INT | 3232–32 | — |
| img_compression | INT | 00–100 | — |
| task_mode | COMBO | auto_from_timeline | 26 options: auto_from_timeline, t2va, i2va, fl2va, ref2va, ref2va_vfx_edit, +20 |
| audio_mode | COMBO | h3_native_generated | 4 options: h3_native_generated, h3_ref2va_audio, h3_custom_audio_drive, external_audio_post |
| prompt_mapping | COMBO | global_plus_local | 3 options: global_plus_local, local_only, global_only |
| performance_profile | COMBO | low_vram_balanced | 14 options: rtx_xx60_safe, rtx_xx70_balanced, rtx_xx80_quality, rtx_xx90_max, low_vram_draft, low_vram_balanced, +8 |
| text_encoder_device | COMBO | gpu_auto | 4 options: gpu_auto, cpu_direct, auto, cpu_safe_12gb |
| seed | INT | 420–18446744073709550000 | — |
| seed_control_after_generate_compat | STRING | fixed | — |
| seed_stride | INT | 10–18446744073709550000 | — |
| steps | INT | 161–100 | — |
| sampler_name | COMBO | euler | 44 options: res_multistep, euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, +38 |
| scheduler | COMBO | simple | 9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3 |
| denoise | FLOAT | 1.000–1 | — |
| shift_video | FLOAT | 12.000.01–100 | — |
| shift_audio | FLOAT | 3.000.01–100 | — |
| turbo_mode | COMBO | off | 3 options: off, early_8_10, ckpt500_6_8 |
| turbo_lora_name | COMBO | 1 options: | |
| turbo_strength | FLOAT | 1.000–4 | — |
| turbo_sampler_mode | COMBO | audio_fixed | 2 options: audio_fixed, res_multistep_stock |
| reference_resize_policy | COMBO | off | 4 options: canvas_crop, canvas_pad, total_pixels, off |
| reference_resize_megapixels | FLOAT | 0.500.1–2 | — |
| reference_resize_filter | COMBO | area | 4 options: area, bilinear, bicubic, nearest-exact |
| acceleration | COMBO | low_vram_auto | 21 options: low_vram_auto, native, h3_sage, h3_sla, sol_low_vram, sol_adaptive_safe, +15 |
| ref_image_size | COMBO | match | 2 options: match, max |
| sol_conditioning | COMBO | exact_kv_and_rows | 2 options: exact_kv_and_rows, exact_kv |
| spectrum_profile | COMBO | low_vram | 5 options: low_vram, quality, aggressive, conservative_3060, conservative_quality |
| vram_clean_before_decode | BOOLEAN | true | — |
| rife_mode | COMBO | off | 3 options: off, rife_48fps, rife_60fps |
| reference_role_1 | COMBO | subject_identity | 5 options: subject_identity, keyframe, composition, style, disabled |
| reference_role_2 | COMBO | subject_identity | 5 options: subject_identity, keyframe, composition, style, disabled |
| reference_role_3 | COMBO | composition | 5 options: subject_identity, keyframe, composition, style, disabled |
| reference_role_4 | COMBO | style | 5 options: subject_identity, keyframe, composition, style, disabled |
| reference_video_role | COMBO | off | 5 options: off, motion_camera, temporal_structure, video_edit, continuation |
| reference_audio_role | COMBO | off | 5 options: off, voice_timbre, rhythm_timing, audio_reuse, sound_reference |
| voice_reference_picture_index | INT | 00–4 | — |
| v2v_guide_mode | COMBO | raw_only | 5 options: , raw_only, raw_pose, raw_depth, raw_depth_pose |
| v2v_source_range_policy | COMBO | timeline_segment | 4 options: , timeline_segment, sequential_requested, repeat_from_offset |
| v2v_source_offset_seconds | FLOAT | 0.000–86400 | — |
| v2v_source_fit | COMBO | canvas_pad | 5 options: , native_adapt, canvas_pad, canvas_crop, stretch |
| v2v_source_end_policy | COMBO | hold_last_for_grid | HOLD LAST FOR GRID requires the complete visible source and freezes only H3's hidden 17k+5 tail (recommended). HOLD LAST VISIBLE also extends a genuinely short source with its final frame and silence; continuity survives but lip-sync cannot continue beyond the real source. ERROR requires real source frames even for the technical tail. |
| v2v_audio_pairing | COMBO | pair_with_source_video | 4 options: , pair_with_source_video, standalone_reference, off |
| flf_join_mode | COMBO | h3_keyframe_cut | 2 options: h3_keyframe_cut, wan_overlap_blend |
| flf_overlap_frames | INT | 00–24 | — |
| flf_continuity_mode | COMBO | stable_keyframes | 3 options: stable_keyframes, native_av_context, longvid_latent_tail_experimental |
| flf_continuity_tail_frames | COMBO | 22 | 3 options: 22, 39, 56 |
| flf_continuity_audio | BOOLEAN | true | — |
| upscale_mode | COMBO | off | 10 options: off, rtx_final, pixel_tiled_low_vram, h3_fast_latent_2pass, h3_pixel_refine, h3_ultimate_tiled, +4 |
| upscale_enabled | BOOLEAN | false | — |
| upscale_width | INT | 1920256–7680 | — |
| upscale_height | INT | 1080256–4320 | — |
| upscale_prompt | STRING | — | |
| upscale_sage | BOOLEAN | true | — |
| upscale_seed_offset | INT | 100000–18446744073709550000 | — |
| wan_upscale_denoise | FLOAT | 0.200–1 | — |
| ltx_seam_safe | BOOLEAN | true | — |
| ltx_detailer_enabled | BOOLEAN | false | — |
| ltx_detailer_lora_name | COMBO | 1 options: | |
| ltx_detailer_strength | FLOAT | 0.600–2 | — |
| ltx_4k_enabled | BOOLEAN | false | — |
| ltx_4k_quality | COMBO | ULTRA | 4 options: ULTRA, HIGH, MEDIUM, LOW |
| ltx_looper_temporal_tile_size | INT | 8024–1000 | — |
| ltx_looper_temporal_overlap | INT | 2416–80 | — |
| ltx_looper_guiding_strength | FLOAT | 1.000–1 | — |
| ltx_looper_overlap_strength | FLOAT | 0.500–1 | — |
| ltx_looper_cond_image_strength | FLOAT | 1.000–1 | — |
| ltx_looper_horizontal_tiles | INT | 11–6 | — |
| ltx_looper_vertical_tiles | INT | 11–6 | — |
| ltx_looper_spatial_overlap | INT | 11–8 | — |
| face_detailer_enabled | BOOLEAN | false | — |
| face_detailer_profile | COMBO | balanced | 6 options: balanced, small_faces, gentle, sam_face_mask, wide_character_12gb, multi_face |
| face_detailer_use_sam_mask | BOOLEAN | false | — |
| h3_upres_model_name | COMBO | Choose an installed H3 3D latent upscaler. Auto uses an installed FP16 checkpoint first; no model is downloaded. | |
| h3_upres_precision | COMBO | fp16 | 3 options: fp16, bf16, fp32 |
| h3_upres_device | COMBO | cuda | 2 options: cuda, cpu |
| h3_upres_keep_models_resident | BOOLEAN | false | — |
| h3_upres_steps | INT | 11–20 | — |
| h3_upres_denoise | FLOAT | 0.200–1 | — |
| h3_upres_sampler | COMBO | sa_solver | 44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38 |
| h3_upres_scheduler | COMBO | simple | 9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3 |
| h3_upres_temporal_chunk | INT | 8517–1700 | — |
| h3_upres_temporal_overlap | INT | 170–850 | — |
| h3_upres_anchor_strength | FLOAT | 0.9990–1 | — |
| h3_upres_tile_width | INT | 86432–4096 | — |
| h3_upres_tile_height | INT | 48032–4096 | — |
| h3_upres_overlap_width | INT | 1280–2048 | — |
| h3_upres_overlap_height | INT | 1280–2048 | — |
| h3_upres_fade_width | INT | 320–2048 | — |
| h3_upres_fade_height | INT | 320–2048 | — |
| h3_upres_min_tile_size | INT | 2560–4096 | — |
| h3_upres_overlap_mode | COMBO | earlier | 2 options: earlier, later |
| h3_upres_overlap_blend | COMBO | linear | 4 options: linear, smoothstep, overwrite, midpoint |
| h3_upres_rtx_enabled | BOOLEAN | false | — |
| h3_upres_rtx_quality | COMBO | ULTRA | 4 options: ULTRA, HIGH, MEDIUM, LOW |
| h3_upres_pixel_groups | INT | 11–16 | — |
| h3_upres_window_frames | INT | 13634–510 | — |
| h3_upres_window_overlap | INT | 220–170 | — |
| h3_upres_pixel_method | COMBO | rtx_vsr | 4 options: rtx_vsr, lanczos-ish bicubic, bilinear, nearest-exact |
| secondary_lora_enabled | BOOLEAN | false | — |
| secondary_lora_name | COMBO | 1 options: | |
| secondary_lora_strength | FLOAT | 0.00-2–2 | — |
| motion_context_window_frames | INT | 36256–362 | Maximum native H3 sample window including the carried AV tail. 362 restores the proven R37/R41/R42 contract. Smaller values are explicit memory trade-offs. |
| h3_exact_profile | COMBO | rtx_xx60_8_12gb_124 | Editable hardware-class baseline: xx60 124f/5.17s, xx70 209f/8.71s, xx80 294f/12.25s, xx90 362f/15.08s. It never overrides the visible window, memory, ClipProj or sampling boxes. |
| h3_exact_chunk_rows | INT | 2048256–65536 | 2048 is the conservative 8–12 GB value. Larger values can be faster but use more activation memory. |
| h3_exact_precision_mode | COMBO | Preserve native | Preserve native keeps INT8/W4A8 checkpoints in their released precision without an implicit conversion. |
| h3_exact_qkv_streaming | COMBO | Auto | Auto preserves the current dense attention backend and adds bounded Q/K/V carriers only where compatible. |
| h3_exact_attention_memory | COMBO | Standard | Lower VRAM performs extra exact attention work. It is a safety fallback, not an approximate cache. |
| h3_clipproj_profile | COMBO | 4b_v3.1 | 4B v3.1 is the default lower-memory conditioning encoder. 8B is larger and heavier. Off uses the connected workflow fallback CLIP. This choice does not replace the MiniMax H3 denoiser model. |
| h3_clipproj_load_mode | COMBO | dynamic | Dynamic uses ComfyUI-managed offload. Streaming lowers the peak further at a speed cost. |
| h3_r40_seed_scout_enabled | BOOLEAN | false | Candidate-scout workflows only. Samples low-resolution candidates before one selected take enters delivery; ordinary generation workflows ignore this field. |
| h3_r40_candidate_count | INT | 31–4 | 1 keeps normal cost; 2-4 create independent cacheable seed candidates. Use 3 for hero-shot scouting, not for every automatic LongVid chunk. |
| h3_r40_seed_stride | INT | 10000031–18446744073709550000 | Candidate N uses base seed + N × stride. 1,000,003 provides widely separated deterministic candidates without a hidden prompt or model override. |
| h3_r40_preview_max_frames | INT | 491–209 | Limits each candidate preview batch only; it never shortens or changes the generated AV latent. |
| h3_r40_sparse_enabled | BOOLEAN | false | Optional current H3SparseAttention. Faster/lower VRAM can change motion, detail or prompt adherence; Off is the quality-safe default. |
| h3_r40_sparse_video_budget | FLOAT | 0.300.05–1 | Fraction of video KV attention retained. 1.0 is dense-equivalent routing; 0.30 is faster but is not lossless. |
| h3_r40_sparse_denser_edges | BOOLEAN | true | Adds attention budget to the first and last sampling steps to protect layout, prompt adherence and final detail. |
| pdd_lora_name | COMBO | Select a native Kijai *_Acc-8Step*_comfy LoRA from models/loras and match FL2VA/Ref2VA to the loaded trunk. Legacy aptech pdd_acc files require their companion loader and are rejected here. | |
| pdd_strength | FLOAT | 1.000–2 | Keep 1.0. Native shape-changing PDD head banks require full trained strength; any different visible Queue-truth value fails clearly instead of being silently replaced. |
| h3_controlnet_enabled | BOOLEAN | false | OFF preserves the standard generation path exactly. ON requires a preprocessed control-video CineLinX module and an installed H3 Fun ControlNet. |
| h3_controlnet_name | COMBO | Select the MiniMax H3 Fun Union checkpoint from models/model_patches. Existing IAMCCS installs in models/controlnet remain compatible. | |
| h3_controlnet_kind | COMBO | pose_dwpose | Declare the already-preprocessed input: DWPose, depth, Canny, HED, MLSD or inpaint. This does not run a hidden preprocessor. |
| h3_controlnet_strength | FLOAT | 1.000–2 | 1.0 is the reference pose-control baseline. Lower values release structure; higher values can over-constrain identity and texture. |
| h3_controlnet_start_percent | FLOAT | 0.000–1 | Sampling interval start. 0.0 applies control from the first denoise step. |
| h3_controlnet_end_percent | FLOAT | 1.000–1 | Sampling interval end. 1.0 applies control throughout; earlier release can preserve more natural detail. |
| h3_controlnet_frame_scope | COMBO | timeline_segment | Timeline segment slices the full source by each Shotboard chunk start. Repeat each chunk restarts from source frame zero. |
| h3_controlnet_end_policy | COMBO | strict_match | Strict match hard-fails before sampling if the source cannot cover the chunk. Hold last repeats the final control frame. |
| fused_turbo_model_name | COMBO | Select the installed MATLOWAI/Kijai fused MiniMax H3 INT8 ConvRot model from diffusion_models. This is a full model, not a LoRA, and is restricted to T2VA. | |
| fused_turbo_sigma_preset | COMBO | 4_step | 4, 6 or 8 exact manual-sigma steps. The Fast Preview button sets Euler, 12/3 shifts, denoise 1 and turns other acceleration LoRAs off. |
| longvid_terminal_endpoint_mode | COMBO | hard_image | HARD IMAGE = legacy exact final AddGuide. LATENT FREE CLOSURE = soft sampled-latent closure. PIANOSEQUENZA LINEAR = upstream-parity direct latent tail with static linear temporal release. PIANOSEQUENZA DRIFT = upstream-parity frozen prefix plus sigma-matched dynamic release. PIANOSEQUENZA NATIVE = upstream-parity native latent-tail keyframe on a fresh video target. PIANOSEQUENZA PHASE = upstream-parity phase-aligned direct latent step guides. PIANOSEQUENZA FROZEN = upstream-parity exact frozen tail as target prefix. PIANOSEQUENZA 2 STAGE = legacy serialized alias for the separate Multi-Stage Spatial toggle. PIANOSEQUENZA HD = self-contained two-stage 0.5x→1.0x progressive solve: every continuation chunk carries the previous native editorial tail, applies sigma-matched DRIFT on the protected prefix, solves LOW, performs the learned H3 latent lift, then completes the SAME sigma schedule at HIGH. The HD engine owns its two-stage contract; the separate Multi-Stage Spatial toggle is ignored while HD is selected. Existing Pianosequenza modes remain unchanged; IAMCCS AudioCon remains authoritative. |
| longvid_pianosequenza_2stage_enabled | BOOLEAN | false | OFF is authoritative one-stage generation. ON enables the IAMCCS progressive spatial path and reveals one Stage Profile dropdown for 2-stage/3-stage and safe/full behaviour. Legacy endpoint value pianosequenza_2stage is migrated to DRIFT but cannot override an explicit OFF toggle. |
| longvid_pianosequenza_stage_count | COMBO | 2_stage | 2_STAGE = classic low→high progressive path. 3_STAGE = low→mid→high progressive spatial solve. |
| longvid_pianosequenza_split_mode | COMBO | auto | AUTO dynamically allocates LOW/MID/HIGH denoising steps from the total schedule. Manual uses the explicit stage step fields below. |
| longvid_pianosequenza_auto_profile | COMBO | balanced | BALANCED ≈ 75/25 total distribution. LOW_AUTHORITY keeps more steps in LOW. HIGH_REFINE gives more to later stages. FAST_SAFE is conservative on high-res. PDD_SAFE and TOETAO_SAFE are tuned for compact step budgets. |
| longvid_pianosequenza_manual_stage1_steps | INT | 40–64 | Manual split only. Stage 1 = LOW-resolution denoising steps. |
| longvid_pianosequenza_manual_stage2_steps | INT | 20–64 | Manual split only. In 2-stage this is HIGH. In 3-stage this is MID. |
| longvid_pianosequenza_manual_stage3_steps | INT | 10–64 | Manual split only. Used only when spatial stage count is 3_STAGE (final HIGH stage). |
| upscale_link_to_native | BOOLEAN | false | Pianosequenza multi-stage only. ON makes the selected legal pair authoritative: the first stage starts at LOW and the same LongVid sample finishes at HIGH. Default pair is 640x384 → 1280x768. It does not enable post-delivery upscale. |
| upscale_link_factor | COMBO | 640x384 -> 1280x768 | Linked LOW → HIGH generation pair. 640x384 → 1280x768 is the default. Legacy 2x/3x/4x saved values are accepted and normalized to the default pair. |
| longvid_pianosequenza_stage_profile | COMBO | 2_stage_full | 2 STAGE FULL = LOW H3 → learned lift → HIGH H3 refine. 2 STAGE SAFE DELIVERY = LOW H3 through the full sigma schedule → learned lift to HIGH, no high-resolution H3 forward. 3 STAGE SAFE MID = LOW H3 → MID H3 → learned lift to HIGH, no final HIGH H3 forward. 3 STAGE FULL = LOW H3 → MID H3 → HIGH H3. |
| longvid_guide_window_policy | COMBO | legacy_auto | LEGACY AUTO preserves existing saves: the old latent-tail continuity selection enables Adaptive windows. STANDARD POSITIONED uses ordinary H3 technical windows with positioned guides and no hidden latent-prefix carry (362f is the normal macro recipe). ADAPTIVE LATENT TAIL enables the narrow high-resolution 209/124-style windows with the configured latent tail. |
| h3_pixel_tiled_method | COMBO | cpu_lanczos | 2 options: cpu_lanczos, model_tiled |
| h3_pixel_tiled_model_name | COMBO | 1 options: | |
| h3_pixel_tiled_tile_size | INT | 512128–2048 | — |
| h3_pixel_tiled_overlap | INT | 640–512 | — |
| h3_continuation_checkpoint | STRING | — | |
| h3_continuation_context_frames | COMBO | 22 | 3 options: 5, 22, 39 |
| h3_continuation_handover_mode | COMBO | terminal | TERMINAL = continue from the actual final AV latent boundary (recommended). AUTO = use saved freeze/lock analysis. MANUAL = deliberately ignore N final source frames. |
| h3_continuation_manual_tail_frames | INT | 00–3400 | — |
| h3_refmod_enabled | BOOLEAN | false | Off preserves the exact former H3 conditioning. On injects one saved provider RefMod into every selected H3 chunk before sampling. |
| h3_refmod_name | STRING | Relative name under models/refmods, without .safetensors. Select from the PRO library or type the provider name. | |
| h3_refmod_strength | FLOAT | 1.000–1 | — |
| h3_refmod_retention | FLOAT | 1.000–1 | — |
| h3_refmod_max_tokens | INT | 00–1048576 | — |
| h3_continuation_enabled | BOOLEAN | false | Opt-in. Only the first queued Shotboard chunk starts from the selected external AV checkpoint; later chunks keep their own selected mode. |
| h3_continuation_save_enabled | BOOLEAN | false | Opt-in export contract for a terminal sampled AV latent. It does not activate loading. |
| h3_faceswap_sam_modelopt | COMBO | Installed SAM3 multiplex checkpoint. Required only when no source mask is connected. | |
| h3_faceswap_birefnet_modelopt | COMBO | BiRefNet model used by FACE SWAP v1 to place both identity views on white before stitching them. | |
| h3_faceswap_mask_promptopt | STRING | head | Tracked identity region. Keep head for hair, ears, jaw and profile stability. |
| h3_faceswap_thresholdopt | FLOAT | 0.300–1 | Foxydits reference threshold for stable single-subject SAM3 tracking. |
| h3_faceswap_objectsopt | STRING | Leave empty for one subject. For multi-person footage, enter the intended tracked-object index explicitly. | |
| h3_faceswap_cleanup_thresholdopt | FLOAT | 0.300–1 | — |
| h3_faceswap_cleanup_shrinkopt | INT | 121–128 | — |
| h3_faceswap_cleanup_min_framesopt | INT | 41–64 | — |
| h3_faceswap_cleanup_edge_growopt | INT | 160–128 | — |
| h3_faceswap_crop_scaleopt | FLOAT | 1.751–4 | — |
| h3_faceswap_crop_megapixelsopt | FLOAT | 0.500–4 | Minimum crop sampling area; 0 preserves the tracked crop size. Editable workload, not a GPU guarantee. |
| h3_faceswap_grow_spatialopt | INT | 360–256 | — |
| h3_faceswap_grow_temporalopt | INT | 10–64 | — |
| h3_faceswap_featheropt | INT | 160–256 | — |
| h3_sla_sparsityopt | FLOAT | 0.850–0.95 | SLA attention only: fraction of key blocks skipped. LightX2V Turbo SLA training recipe uses 0.85. |
| h3_sla_dense_last_stepsopt | INT | 00–20 | SLA attention only: optional final dense steps for an explicit quality comparison. |
| h3_advisor_stateopt | STRING | {} | Per-mode local advisor state. Proposals apply only after acceptance. |
| seed_policyopt | COMBO | fixed_per_generation | Fixed/random per generation: one seed for all chunks. Fixed per chunk: base + index × stride. Random per chunk: independent seeds derived from the randomized queue seed. Random policies choose a new base when you press Queue in Settings Pro; API clients must supply their base seed. |
| h3_editor_enabledopt | BOOLEAN | false | OFF by default: all technical chunks are still joined into the normal final master and exported directly. ON requests the optional Shotboard Video Editor branch after master assembly. |
| keyframe_joint_latent_newopt | BOOLEAN | true | Keyframe Joint only: experimental B1 MASKED+EXACT windows with positioned destination guides. No LatentGoAhead dependency. OFF preserves the legacy single-sample path. |
| h3_continuation_visual_handoveropt | COMBO | auto_safe | AUTO SAFE: if Shotboard first/last are the same still, use it only as a Qwen identity reference and let saved AV own the temporal opening; a distinct last image remains a future destination. LATENT ONLY: saved AV + prompt only. SHOTBOARD DESTINATION: legacy behavior, explicitly force the board endpoint. |
| h3_continuation_run_and_gun_enabledopt | BOOLEAN | true | When continuing from a checkpoint with IAMCCS chain metadata, also build one cumulative A+B(+C...) master from the original segment files. Individual segment files stay available for the Video Editor. |
| h3_continuation_run_and_gun_joinopt | COMBO | soft_av | SOFT AV (recommended) uses the incoming generation's hidden, time-corresponding AV context to soften the outgoing boundary without changing latent continuation or programme duration. DIRECT preserves the exact hard join. |
| h3_continuation_soft_video_framesopt | INT | 40–16 | Recommended: 4. Delivery-only context-aligned blend; 0 disables video blending. |
| h3_continuation_soft_video_curveopt | COMBO | smoothstep | SMOOTHSTEP is the recommended short dissolve. LINEAR is available for a strictly uniform ramp. |
| h3_continuation_soft_audio_msopt | FLOAT | 150–100 | Recommended: 15 ms equal-power de-click against the matching hidden audio context. Keep it short for dialogue and lipsync. |
| h3_faceswap_generate_new_audioopt | BOOLEAN | false | OFF (recommended): use source_video audio as Ref2VA lip timing and preserve that exact audio in the output. ON: ignore source audio for conditioning/output and generate a new native H3 audio stream. |
| h3_controlnet_use_video_audioopt | BOOLEAN | false | With CUSTOM AUDIO DRIVE: ON uses the source control video's audio. OFF uses the Shotboard AudioBoard. If the video has no audio, AudioBoard remains the fallback. Native Generated ignores both routes. |
| h3_controlnet_dwpose_torch_gpuopt | BOOLEAN | false | ON: when Cine H3 Fun Control Input preprocesses source_video as DWPose, use TorchScript YOLOX + DWPose on the ComfyUI CUDA device. OFF: preserve the existing ControlNet Aux automatic route. Already-preprocessed control_video bypasses this selection. |
| h3_controlnet_preprocess_resolutionopt | COMBO | node_default | Choose the ControlNet Aux preprocessing resolution for source_video. NODE DEFAULT preserves the Cine H3 Fun Control Input widget for older workflows. Already-preprocessed control_video bypasses this choice. |
| h3_auto_extend_modeopt | COMBO | off | When a non-LongVid H3 shot exceeds one native 17k+5 window, automatically compile it to a motion-context auto-chain so the action can continue past 15 seconds without manually switching mode. |
| h3_auto_extend_continue_audioopt | BOOLEAN | true | AUTO EXTEND only. ON keeps audio continuity through the motion-context chain. OFF continues visually while leaving audio on the standard route. |
| h3_fast_latent_temporal_chunkopt | COMBO | AUTO | FAST LATENT 2-PASS learned 3D upscaler temporal chunk only. This does NOT control the H3 Stage-2 refine sampler. |
| h3_fast_latent_stage2_windowopt | COMBO | AUTO | Temporal window for the H3 Stage-2 refine. AUTO uses a conservative GPU-aware window; RTX 3060/12GB at ~1.4MP resolves to 56 frames. FULL reproduces the old all-frames-at-once sampler and can OOM on 12GB. |
| h3_fast_latent_stage2_overlapopt | COMBO | AUTO | Overlap between Stage-2 temporal refine windows. AUTO uses 22 frames for continuity when possible; 5 reduces repeated work. |
| extended_av_profileopt | COMBO | safe_8_12gb | 5 options: safe_8_12gb, balanced_12_16gb, quality_20_24gb, max_32gb_plus, custom |
| extended_av_pin_modeopt | COMBO | masked | 2 options: masked, both |
| extended_av_mask_profileopt | COMBO | exact | 3 options: exact, runway, soft |
| extended_av_custom_window_framesopt | COMBO | 192 | 7 options: 39, 90, 141, 192, 243, 294, +1 |
| extended_av_custom_overlap_framesopt | COMBO | 39 | 6 options: 39, 90, 141, 192, 243, 294 |
| extended_av_handover_modeopt | COMBO | auto | 3 options: auto, same_time_equal_power, direct |
| extended_av_level_lock_modeopt | COMBO | auto | 2 options: auto, off |
| extended_av_joint_refine_modeopt | COMBO | off | 1 options: off |
| extended_av_joint_window_framesopt | COMBO | 192 | 7 options: 39, 90, 141, 192, 243, 294, +1 |
| extended_av_joint_window_overlapopt | COMBO | 39 | 6 options: 39, 90, 141, 192, 243, 294 |
| extended_av_soft_audio_handover_msopt | FLOAT | 155–60 | Audio only. Length of the hidden-child equal-power handover at EXTEND joins. 15 ms is the proven baseline. Video is never changed. |
| extended_av_boundary_polish_msopt | FLOAT | 3.00–12 | Audio only. Duration-preserving endpoint correction after the hidden Soft-AV handover. 0 disables it; 2-5 ms is the recommended tuning range. |
| extended_av_boundary_polish_strengthopt | FLOAT | 1.000–1 | Audio only. 1.0 exactly matches the first visible child sample to the preceding handover endpoint, then the correction decays to zero. Does not alter video or duration. |
| h3_fast_latent_stage2_guidesopt | COMBO | soft | Soft preserves target-resolution guide conditioning in refinement windows. Masked additionally pins image guides to their containing video latent token; audio is unchanged. Fast Latent 2-Pass only. |
| h3_reference_image_sourceopt | COMBO | auto_prompter_then_direct | AUTO first uses an image attached to IAMCCS Prompter, then the direct image socket. PROMPTER REF2V requires an attached visual that is injected with the prompt. DIRECT REFERENCE uses the dedicated Load Image socket. |
| h3_reference_image_presetopt | COMBO | single_image | Single production still or one identity-consistent character sheet canvas. |
| h3_reference_image_detail_enabledopt | BOOLEAN | false | OFF is the native Sato output. ON applies the selected detail model and returns to the exact native canvas. |
| h3_reference_image_detail_modelopt | COMBO | None | Optional local upscale/detail model. No model is downloaded and None keeps the branch unchanged. |
| h3_reference_image_detail_strengthopt | FLOAT | 0.600–1 | Blend of the same-size micro-detail pass. Lower values preserve the native image more strongly. |
| face_detailer_modeopt | COMBO | auto_from_profile | AUTO FROM PROFILE keeps legacy behavior: profile MULTI FACE selects sequential multi-subject refinement, every other profile selects one subject. SINGLE always refines one tracked subject. MULTI FACE refines up to the configured number of distinct tracked subjects sequentially, so VRAM peak remains close to one face pass. |
| face_detailer_max_facesopt | INT | 41–8 | Upper bound only. If the shot contains fewer distinct faces, only the detected subjects are refined; a one-face shot performs one pass. |
| face_detailer_min_coverageopt | FLOAT | 0.100–1 | Multi Face only. Skip incidental tracks detected in less than this fraction of the source frames. 0.10 is a conservative crowd default. |
| face_detailer_canvas_widthopt | INT | 512256–1536 | Crop render width for the face pass. Larger canvases can recover more detail but increase VRAM and time. |
| face_detailer_canvas_heightopt | INT | 512256–1536 | Crop render height for the face pass. Keep width and height equal unless you have a deliberate reason not to. |
| face_detailer_crop_factoropt | FLOAT | 2.501–6 | How much context around the face is included. Lower values make the face larger inside the crop; higher values preserve more head/background context. |
| face_detailer_confidenceopt | FLOAT | 0.350.01–1 | Lower values catch smaller or softer faces; higher values are stricter and can reduce false positives. |
| face_detailer_stepsopt | INT | 41–100 | Sampling steps for each face-refine crop. Higher values can improve detail at the cost of time. |
| face_detailer_denoiseopt | FLOAT | 0.200–1 | How freely the face pass may rebuild the crop. Higher values increase detail and change; lower values preserve the native take more strongly. |
| face_detailer_window_framesopt | INT | 7334–510 | Temporal chunk size for the face pass. Larger windows can improve coherence but raise memory and time. |
| face_detailer_window_overlapopt | INT | 220–170 | Temporal overlap between face chunks. Must stay smaller than the window size. |
| face_detailer_strength_small_faceopt | FLOAT | 1.000–1 | Relative denoise multiplier for smaller detected faces. Keep near 1.0 when distant faces need help. |
| face_detailer_strength_large_faceopt | FLOAT | 0.350–1 | Relative denoise multiplier for larger detected faces. Lower values preserve identity more strongly on close faces. |
| face_detailer_blendopt | FLOAT | 0.700–1 | How much of the refined face is pasted back. Higher values reveal more generated detail; lower values keep more of the native pixels. |
| face_detailer_paste_regionopt | COMBO | face_only | FACE ONLY pastes only the refined face area. CROP pastes the full refined crop back to the frame. |
| face_detailer_featheropt | INT | 60–128 | Blend softness around the pasted region. Slight feathering helps avoid visible seams. |
| longvid_continuous_engineopt | COMBO | pianosequenza_1_native_tail | Native TimelineDirector with selectable 22/39/56 incoming-owned overlap; Extended AV uses its own masked pin/trim backend and VRAM preset; Evolving uses timed Shotboard prompt beats on native AV tail; DRIFT/PHASE experimental terminal variants. Native Delivery is recommended for validation. |
| h3_reference_image_vram_profileopt | COMBO | up_to_12gb_safe | Native sampling cap before delivery resize: <8GB 0.55MP; 8-12GB safe 1.05MP; 12GB quality 1.50MP (test for OOM); up to 24GB 4MP; over 24GB 8.5MP (native 4K). CUSTOM uses the cap on SatoDive. Actual capacity also depends on model quantization and reference count. |
| h3_veda_enabledopt | BOOLEAN | false | Experimental on RTX 30: kernel self-test required. Applied after LoRAs and compatible memory patches; does not change sampling steps. |
| h3_veda_predictoropt | COMBO | auto | 3 options: auto, minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors, minimax_h3_r2va_veda_preview_fp8.safetensors |
| h3_veda_verboseopt | BOOLEAN | false | — |
| h3_swap_lora_nameopt | COMBO | 1 options: | |
| h3_swap_lora_strengthopt | FLOAT | 1.000–2 | — |
| h3_swap_realism_lora_nameopt | COMBO | 1 options: | |
| h3_swap_realism_strengthopt | FLOAT | 1.000–2 | — |
| h3_swap_refine_seedopt | INT | 7770–18446744073709550000 | — |
| h3_upres_seedopt | INT | -1-1–18446744073709550000 | Fixed Stage-2 seed plus segment stride. -1 preserves the legacy seed/offset or SwapChar refine seed. H3 Detail preset selects 777. |
| h3_vfx_lora_nameopt | COMBO | 1 options: | |
| h3_vfx_lora_strengthopt | FLOAT | 1.000–2 | — |
| h3_edit_video_sourceopt | COMBO | auto | 3 options: auto, prompter, backend |
| h3_attention_backendopt | COMBO | from_acceleration | Independent of sampling steps and Turbo LoRA. From acceleration preserves existing workflows. FastH3, SLA, Fused and Progressive manage their own attention. Sol and SLA/Fused cannot be combined with VEDA. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| cine_linx | IAMCCS_SUPERNODE_LINX | — |